Someone listed their place in San Francisco this week for just under $3 million, and right there in the listing, they say they'll take Anthropic or OpenAI stock as payment.

The final boss of San Francisco real estate is a 28-year-old who'll let you buy his apartment with pre-AGI equity. I laughed at this then remembered that Anthropic just raised $65 billion at a $965 billion valuation, sailed past OpenAI, and the house thing stopped being such a joke.

Last week I made the case that the thing that matters most with all this isn't which AI you use, it's whether you're actually using it every day on real work. Keep that in your back pocket, because this week kind of proves it.

First, the money, because it's hard to even picture. Anthropic is now worth around $965 billion, more than OpenAI, and bringing in roughly $47 billion a year. A year ago a number like that would've sounded made up. But the giant number isn't really the part that changes your week.

The same day, they put out a new version of Claude (Opus 4.8, if you're keeping count). The scores nudged up a bit, nothing dramatic. What people care about is the stuff around it.

There's a little dial now that lets you tell it how hard to think, from "quick and cheap" to "really chew on this." There's a faster, cheaper setting. There's a mode where it splits a big job across dozens of copies of itself, all working at once. And there's another change I think is the most interesting one. It's now a lot better at telling you when it's not sure, instead of confidently handing you something that's wrong.

There's a little dial now that lets you tell it how hard to think, from "quick and cheap" to "really chew on this." There's a faster, cheaper setting. There's a mode where it splits a big job across dozens of copies of itself, all working at once. And there's another change I think is the most interesting one. It's now a lot better at telling you when it's not sure, instead of confidently handing you something that's wrong.

Reactions were all over the map.

Dan Shipper, who runs Every and tests this stuff obsessively, loved it and said it's good enough they may as well have skipped ahead and called it Opus 5. Anthropic's own CTO said the "tells you when it's unsure" part matters to him more than any test score. Then you scroll over to where the heavy users hang out, and they're fuming. They loved the old version, they reckon the new one burns through their usage, and they're roasting it for sounding fake-humble (it keeps opening replies with "let me be honest…"). The best example going around is a guy who asked it for a quick review, and Claude spun up 45 copies of itself and torched his entire usage allowance in 43 minutes.

One reply called it "the new YOLO mode for when you've got usage left before the weekly reset."

Meanwhile, the most-shared setup guide for Opus 4.8 going round this week barely talks about the flashy stuff. It spends most of its time on that little dial, and the big tip is that most people will just leave it cranked all the way up and never touch it. The actual trick is turning it down for the easy stuff and only leaning on it for the hard stuff. Which, from the other direction, is exactly what nuked that guy's afternoon. The AI can't sort that out for you. Same with the "tells you when it's unsure" bit. It only helps if you're the kind of person who reads the part where it says so.

Funny thing is, Anthropic's own CEO more or less called this almost a year ago, before any of it existed. On a podcast last summer he said AI would mess up way less often than people, but in weirder, sneakier ways, because it sounds just as sure of itself when it's wrong as when it's right. The fix, he said, was teaching it to double-check its own work. That's basically what this update is. The hard thing he described is now a setting you can switch on.

And the people who actually do this work for a living keep landing in the same spot. Duolingo's CEO spent this spring walking back his big "AI-first" announcement, now openly admitting the stuff still spits out a lot of junk. A founder who's about as pro-AI as they come told his team a line I liked: AI should get you to great work faster, not drop you straight into mediocre. And Klaudia, a team manager at Ocado who just finished our 15 Days of AI for Teams, said it best. The win for her wasn't AI doing her job. It was setting it up to handle the boring lifting so she could put her energy into the part that's actually hers.

So that's the week. A near-trillion-dollar price tag, a new Claude half the internet loves and the other half wants to undo, a house that'll take your startup shares, and a pile of smart people agreeing the AI was never the hard part. There'll be a slightly better one next month, there always is. The thing I'm watching is whether "when do I lean on this, when do I check it, and what's even worth making" starts to feel like the real skill. For most teams, not yet.